This week I have been continuing my ongoing research and diving into the uncanny, utilised the latest advanced methods for image generation and utilised a workaround that uses optimization techniques typically used in LLM's to be able to run thelarge DEV-Flux model on my gaming laptop with limited gpu space.
Through this praactice i started exploring ideas of using these assets in real sculptural pieces, animating them etc.
I continued my reading of Communicology by Vilem Flusser and read Art in the age of mechanical reproduction by .....{insert author here}.... I also watched Synechdoche new york; which famously uses the ideas of jean beaudrillard as a key part of the narrative;
these pieces have been enhancing my media literacy and understandings of interpretations of impacts of the systems and the current prevalent reward mechanisms and content based feed system or ideology / behavioural system.
FLUX - GGUFF Models and learnt how to use comfyui - specifically youtube tutorials and documentation.
I've found a way to speed up my data collection project using a Python script download_reels.py) that uses Instaloader. I'm now able to collect Instagram Reels and store them locally. This new method allows me to keep a list of good videos on my mobile phone, send them to myself, and clear the list each time, saving me from downloading individual files.
Insert Instaloader Github Project Link
COMFY-UI - FLUXX - GGUFF
This week i began researching how to use the latest ai model FLUX; I had been seeing some fairly wild outcomes across social media.
I began by watching several youtube videos on this topic and realised I had to install comfy ui; previously I had used automatic1111 running sdxl locally with some additional more advanced techniques such as inpainting addons etc but had found the node based input system of comfyui to be a bit daunting.
After installing Comfyui I had to find a workaround for running the latest flux model on low vram as the recommended minimum is 12gb and my laptop only as 6gb. I found someone had made use of an optimisation or memory caching technique typically used for languae models on lower end devices that had been adapted to work with the flux image generation system. After downloading this I began to test out the two different models. Flux Dev and Flux Schnell. Schnell is a model that has been trained from the larger dataset on its output images and therefor is more condensed and faster. However I was seeing better more coherent results from the larger DEV model.
Below are the comparisons between the two models of some generations, i believe you will see what I mean.